McAllister
Blair

Built a 6-degree-of-freedom follower robot arm from hardware assembly through deployment of a trained imitation learning policy. The final system autonomously picks up a block and places it into a bowl.

• Wrote a custom PS4 teleoperation script for collecting demonstrations
• Recorded 50+ episodes and trained an ACT policy on the collected trajectories
• Deployed the trained policy back onto the physical robot and validated it on real hardware

This project gave me hands-on experience debugging a full robotics stack across hardware, webcam input, controller mapping, data collection, and model training on Ubuntu.

Hands-on autonomous vehicles course: building, training, and deploying a full autonomy stack on a physical RC car platform, from data collection through on-vehicle inference.

• Deployed and iterated a full end-to-end behavioral cloning pipeline (DonkeyCar) — driving data collection, model training, and autonomous laps on real hardware
• Ran training workflows on UCSD's DSMLP GPU cluster and deployed trained models to a Raspberry Pi 5 onboard the vehicle
• Debugged the full hardware/software stack: camera input, actuation, calibration, and the train-deploy loop between the cluster and the car

Full write-up, code, and results will be posted here at the end of the course.

Automated R-peak detection on live ECG (click to enlarge)

AD620-based analog front-end (click to enlarge)

Built a complete, end-to-end hardware and software pipeline to capture, digitally filter, and analyze live human biometrics using off-the-shelf components and Python. The front-end was designed with EEG acquisition as the long-term goal and was first validated on ECG, where the stronger, well-characterized signal made it possible to verify the full analog and digital chain.

• Analog Hardware: Designed and built a physical amplifier circuit on a breadboard to extract microscopic biological signals while filtering out ambient electrical noise.
• Digital Telemetry: Engineered a safe hardware bridge to an Arduino, capturing real-time physiological data at a ~190Hz effective sampling rate.
• Data Pipeline: Developed a custom Python script using PySerial to ingest the live hardware data stream directly into a standard data science environment.
• Signal Processing: Applied clinical-grade digital bandpass filters (MNE-Python) to remove baseline wander and residual interference from the raw data.
• Algorithmic Validation: Wrote an automated peak-detection algorithm (SciPy) with a strict voltage floor to accurately isolate heartbeats and calculate a live resting heart rate.

Same features, same model — only the split differs.

A cross-subject EEG decoder for Parkinson's medication state — and an investigation into why its first result was wrong. The original pipeline reported 0.796 ROC-AUC from a random train/test split that leaked subject identity. Rebuilt with leave-one-subject-out evaluation and confound-resistant features, it holds at 0.578 across held-out patients, validated against a permutation null

• Ingested, filtered, and epoched raw clinical BDF datasets from OpenNeuro using MNE-BIDS, extracting 48 features across 7 sensorimotor channels.
• Replaced absolute band power with relative power and normalized the PAC modulation index, closing two routes by which per-session recording gain — confounded with the ON/OFF label — leaked into the classifier.
• Found that the two highest-scoring subjects were the only two flagged in the dataset metadata for a preprocessing mismatch between sessions; excluding them improved 8 of the remaining 13.
• Validated against a 100-permutation shuffled-label null (0.500 ± 0.012, p = 0.0099).